Rejectable deep differential dynamic programming for real-time integrated generation dispatch and control of micro-grids

Abstract

With the application of renewable energy and distributed power generation in micro-grids, conventional artificial intelligent control strategies have shown deficiencies for the frequency control and economic dispatch of micro-grids. Conventional deep learning controllers could provide outputs although when the predicted probability is not high, which will lead to micro-grid system divergence. This paper proposes a rejectable deep differential dynamic programming for the real-time integrated generation dispatch and control of micro-grids. The rejectable deep differential dynamic programming can provide an action from an analytic control algorithm when the predicted probability is not high enough. The deep differential dynamic programming contains four deep neural networks, i.e., “deep differential prediction network”, “deep differential evaluation network 1”, “deep differential evaluation network 2” and “deep differential execution network”. To verify the feasibility and effectiveness of the proposed rejectable deep differential dynamic programming, a total of 25 combined conventional optimization and control algorithms are compared under a micro-grid based on Hainan Power Grid. The numeric simulation results show that the proposed approach can obtain high control performance for the real-time integrated generation dispatch and control framework, which can replace the conventional combined “economic dispatch + automatic generation control + droop control” framework of micro-grids.